stemdeckapp/stemdeck v0.12.0

0.12.0

on GitHub

xattr -dr com.apple.quarantine /Applications/StemDeck.app

What's new in 0.12.0

Lead/backing vocal split. An on-demand action that splits the Vocals stem into separate Lead Vocals and Backing Vocals lanes, plus two bug fixes to the stems relocation feature and faster first-run setup on desktop.

StemDeck is still alpha. A leading 0. means no stability promise.

Lead/backing vocal split

• New per-track action, not part of every separation. On the Extract bar, the Vocals stem now has an "All / Lead + Backing" choice. Picking Lead + Backing runs a second, focused pass over the isolated vocals stem once the main separation finishes, splitting it into Lead Vocals and Backing Vocals lanes in the mixer, in place of the single Vocals lane. The original Vocals file is never modified, just supplemented.
• Runs on-demand, not upfront, since most tracks don't need it: choosing it queues the split right after the base separation completes, whether that job is in the foreground or already running in the background.
• Model: UVR-MDX-NET Karaoke 2, via the open-source audio-separator library, trained as part of the Ultimate Vocal Remover project (credit in the README, license details in docs/models.md). A community Roformer-based alternative was evaluated and rejected: it turned out to have no usable license (distributed only through a supporter paywall, no LICENSE file, its public mirror already taken down).
• The model (roughly 50 MB) downloads once and is cached alongside the Demucs and beat-detection models.

Stems relocation fixes

• Your library metadata now moves with your stems. Favorites, folder layout, and per-track mixer state used to stay behind in ~/Documents/StemDeck when you relocated your stems folder via Settings. That file now lives inside the stems folder itself, so a relocation carries it along automatically.
• A failed relocation no longer reports success. Moving your stems folder writes the new location to a settings file as a separate step from the physical move; if that write silently failed, Settings would keep showing the old location after a restart with no indication anything went wrong. StemDeck now checks that the write actually landed and tells you plainly if it didn't, instead of a false "done."

Faster first-run setup (desktop)

• All three ML models now download during setup, not on first use. Demucs, the beat-detection model, and the new karaoke model used to each download the first time a job actually needed them, adding a delay mid-pipeline. All three now download during the first-boot setup wizard instead, so your first real import doesn't pay for any of them.

Notes on the beat detection model

Unchanged, repeated here for anyone upgrading from an older release:

• The first time you use the click track, StemDeck downloads a 77 MB beat-detection model (MIT licensed, code and weights). It is cached alongside the Demucs models, so this happens once.
• If that download is unavailable, for example on a machine with no internet, detection falls back to the classic librosa tracker automatically. The click still works; it is just more likely to pick half or double tempo on fast material, which /2 and x2 correct.
• Known limitation: tracks whose time signature changes bar to bar, such as odd-time progressive material, get a usable steady pulse but the bar accents will not follow the written meter. No current beat tracker handles that automatically. The grid editor is the way to fix it by hand.
• Tracks with no drums fall back to detecting on the full mix and are less reliable.